Gemini (Search / Web / Analyze / Summarize)

Unified Gemini CLI interface. Leverages Gemini's built-in tools (google_web_search, web_fetch, @file) — all at zero API cost (subscription-based).

Tool

python3 .claude/skills/gemini/scripts/gemini_ctl.py <command> [args]

Commands

# Web search (Gemini uses google_web_search internally)
python3 .claude/skills/gemini/scripts/gemini_ctl.py search "latest developments in AI agents"
python3 .claude/skills/gemini/scripts/gemini_ctl.py search "OpenAI o3 发布" --lang zh

# Read and process a URL (Gemini uses web_fetch internally)
python3 .claude/skills/gemini/scripts/gemini_ctl.py web "https://example.com/article"
python3 .claude/skills/gemini/scripts/gemini_ctl.py web "https://example.com/api-docs" --prompt "Extract authentication methods and rate limits"

# Analyze a file (image, PDF, code, text — any format)
python3 .claude/skills/gemini/scripts/gemini_ctl.py analyze /path/to/document.pdf
python3 .claude/skills/gemini/scripts/gemini_ctl.py analyze /path/to/image.png --prompt "Describe what's in this image"
python3 .claude/skills/gemini/scripts/gemini_ctl.py analyze /path/to/data.csv --prompt "What trends do you see?"

# Summarize long content (auto-detects URL vs file)
python3 .claude/skills/gemini/scripts/gemini_ctl.py summarize /path/to/long-report.pdf
python3 .claude/skills/gemini/scripts/gemini_ctl.py summarize "https://example.com/long-article" --max-chars 1000

# Check Gemini CLI availability
python3 .claude/skills/gemini/scripts/gemini_ctl.py status

When to Use

Scenario	Command	Why Gemini?
Web search (any language)	search	Google Search Grounding — default search, complements Brave
User shares a URL to read/summarize	web	Keeps large page content out of CC context
User asks about an uploaded image	analyze	Gemini handles images via @file, free
User asks about an uploaded document	analyze	Document stays with Gemini, CC context clean
Long article/transcript for summary	summarize	Handles large content without CC context bloat

Search routing: Use search as the default for web search. Fall back to Brave Search MCP or CC WebSearch when Gemini CLI is unavailable or when you need structured result metadata (URLs only, count-based queries).

Do NOT use when:

• User needs CC tool capabilities (file editing, git, bash) — use CC directly
• Simple question CC can answer from knowledge — no need to delegate
• Need precise structured data extraction — CC is better at following exact output formats

How It Works

Gemini CLI runs as a stateless subprocess with built-in tools:

• google_web_search — Google Search Grounding (real-time web search)
• web_fetch — fetch and process web pages
• @/path/to/file — inject local files into context (no upload cost)

Each invocation is independent (no session persistence in pipe mode). The script constructs a targeted prompt, pipes it to gemini via stdin, and returns plain text output.

Context Strategy

• CC never sees raw content — Gemini processes it and returns a concise result
• File analysis: File stays on disk, Gemini reads via @file, CC gets the summary
• Web content: Gemini fetches and processes, CC gets the synthesis
• This keeps CC's context window focused on the conversation

Options

Option	Commands	Default	Description
--prompt TEXT	analyze, web, summarize	(varies)	Custom instruction
--lang zh|en|auto	search	auto	Output language preference
--max-chars N	summarize	1500	Target output length
--model MODEL	all	(per-command)	Override Gemini model
--timeout N	all	varies	Timeout in seconds

Model Defaults

Command	Default Model	Rationale
search	gemini-3-flash-preview	Fast, tool-heavy search tasks
web	gemini-3-flash-preview	URL fetching + extraction
analyze	gemini-3.1-pro-preview	Deep reasoning for document analysis
summarize	gemini-3.1-pro-preview	Quality synthesis for long content

Override with --model flag when needed. All models use subscription (zero API cost).

File Locations

Documents uploaded via Feishu are stored at:

data/files/<session_key>/<timestamp>_<filename>

Use file_store metadata (shown in session context) to find the file path.

Fallback

If Gemini CLI is unavailable or fails:

• Search: Use CC's WebSearch tool or Brave Search MCP
• Web: Use CC's WebFetch tool
• Analyze (PDF): Read with CC's Read tool (pages parameter, max 20 pages/request)
• Analyze (image): CC has native vision via Read tool
• Summarize: Read the file directly into CC context (watch for length)